VLA-JEPA MolmoAct2 BimanualYAM (3-cam, unfrozen Qwen)
Fine-tuned VLA-JEPA on the MolmoAct2-BimanualYAM-Dataset with 3 cameras (top, left, right) and an unfrozen Qwen3-VL-2B backbone.
Checkpoint details
| Setting | Value |
|---|---|
| Training step | 050000 |
| Cameras | top, left, right |
| Action dim | 14 (joint positions + grippers) |
| Chunk size | 30 (@ 30 Hz) |
| Qwen | Qwen/Qwen3-VL-2B-Instruct (unfrozen) |
| Base model | lerobot/VLA-JEPA-Pretrain |
| Gripper range | [0, 1] (continuous, not binarized) |
Usage (LeRobot)
from lerobot.policies.vla_jepa.modeling_vla_jepa import VLAJEPAPolicy
policy = VLAJEPAPolicy.from_pretrained("Jiafei1224/VLA-JEPA-MolmoAct2YAMData")
Usage (MolmoAct2 sim eval server)
See the MolmoAct2 sim_eval docs — serve via an /act HTTP endpoint with all three camera views.
Eval (step 050000)
- Open-loop: MSE 0.127, MAE 0.197 (50 held-out episodes)
- Closed-loop sim: 0% on
BimanualYAMPutEverythingInBox-v1(infra OK, task not solved yet)